Mac Studio M5 Ultra: 512GB of unified memory on the desktop

Apple has just introduced the new Mac Studio with M5 Max and M5 Ultra chips, and the top configuration — 512GB of unified memory, arriving in late October — is a turning point for anyone who wants to run frontier-class models locally, without a cloud and without depending on an API. For the first time, a consumer desktop has enough memory to load MoE models of several hundred billion parameters, at 4 bits, with a large context. In this article: what changes, the speed math, the cost comparison with NVIDIA's DGX Station, and when the Mac is really the right choice. Primary sources: Apple's official announcement (Apple Newsroom) and the spec sheet.

The new Mac Studio in brief

The Mac Studio comes in two chip families:

• M5 Max: 18-core CPU, up to a 40-core GPU with Neural Accelerators, up to 128GB of unified memory. From $2,499.
• M5 Ultra: up to a 36-core CPU, up to an 80-core GPU, up to 512GB of unified memory at 1.2 TB/s. From $5,499; the 512GB configuration is $9,999 and lands in late October, after the rest of the line goes on sale on September 22 (pre-orders opened August 25).

Apple measures the generational jump at up to 4.3x faster on inference workloads, thanks to Neural Accelerators built directly into the GPU cores. Wi-Fi 7 and Bluetooth 6 debut on the Mac Studio.

M5 Ultra: the first quad-die chip

The M5 Ultra is Apple's first chip built on four dies: two M5 Max fused by UltraFusion technology, presented to the system as a single processor. The practical consequence for local AI is memory: up to 512GB unified, shared across CPU, GPU and Neural Accelerators, with 1.2 TB/s of bandwidth. No data copies between CPU and VRAM: model, context and RAG data all live in the same pool. It is the unified-memory principle taken to its most extreme limit.

What fits: MoE from 100 to 480B

At 4 bits, every billion of parameters weighs about 0.5GB. The arithmetic:

• 100B total (MoE): ~50GB of weights → huge headroom for long context and RAG.
• 200B: ~100GB.
• 400B: ~200GB → room left for contexts of several hundred thousand tokens.
• 480B: ~240GB → it fits, but the context needs managing.
• 560B (the edge case): ~280GB at 4 bits → fits in 512GB with headroom; not at 8 bits.

A concrete model in this class is Ling 3.1 Flash from inclusionAI: 560B total, 25B active per token, up to a 1M-token context, MIT license. At 4 bits it fits entirely in a 512GB Mac Studio — exactly the kind of MoE model that would need a GPU cluster on conventional hardware.

At what speed: the bandwidth math

In MoE models the decode speed is dominated by bandwidth: the bottleneck is the memory read per token. Orders of magnitude (physics estimates, not official benchmarks):

• MoE with 25B active at 4 bits (~12.5GB per token): 1.2 TB/s ÷ 12.5GB → about fifty tokens per second in theory, twenty to thirty in practice with software overhead.
• 200B dense at 4 bits (~100GB per token): a dozen tokens per second in theory, five to eight in practice.

Apple claims up to 4.3x faster than the previous generation. And if you need more still: a cluster of four Mac Studios, linked over Thunderbolt 5 with RDMA, reaches up to 3x the inference speed of a single unit.

Mac Studio 512GB vs DGX Station: the cost comparison

NVIDIA's DGX Station (GB300 Grace Blackwell Ultra) carries up to 748GB of coherent memory, up to 20 petaflops in FP4 and a 72-core Grace CPU linked via NVLink-C2C: announced at GTC in March 2025 (product page, official press release), with a Windows version announced on June 1, 2026, due at the end of the year. The price is in the order of $99,000: roughly ten times the price of the $9,999 512GB Mac Studio.

In usable memory the two systems are the same class (512GB versus 748GB), but the DGX Station has far more raw compute in petaflops. The Mac wins on price, power draw, silence and unified simplicity — and on scaling: four units clustered means roughly 2TB of memory for under $40,000, in a range one professional or a small technical team can actually reach.

When the Mac is the right choice

• You are a desktop user, not a server: you want a 100-560B MoE at 4 bits with long context, without running a rack, data-center noise or a rental contract. The 512GB Mac Studio is the only desktop that does it alone.
• The budget is under $10-20k and you count the cost per token as zero: no API calls, no subscriptions, the model is yours.
• You want predictable vertical scaling: two, three, four units over RDMA, without changing software ecosystems.

The DGX Station remains the choice when you need compute peaks (huge prefills, FP4, the full CUDA ecosystem) or 748GB of coherent memory on a single machine.

In practice

1. Local-AI runtimes (Ollama, LM Studio, mlx-lm) run natively on Apple Silicon: install the runtime, download the model, run it.
2. Start with a 4-bit quantization (GGUF) of the model you care about and do the math: weights + expected context + margin < 512GB.
3. If the model is MoE with few active parameters (Ling 3.1 Flash-type or large sparse MoE), expect conversational speeds. If it is dense and heavy, accept the trade-off: the memory is there, speed depends on bandwidth.
4. Keep an eye on the 512GB configuration: that is the one arriving in late October. The rest of the line is already orderable.

Article generated with AI